Learning from label proportions in brain-computer interfaces: Online unsupervised learning with guarantees

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Traditional brain-computer interfaces (BCIs) require extensive calibration data for new users.
• Existing methods for reducing calibration time, like classifier transfer or unsupervised adaptation, lack theoretical guarantees.
• Event-related potential (ERP) based BCIs are a common but calibration-intensive technology.

Purpose of the Study:

• To develop a reliable, unsupervised, and calibrationless decoding method for ERP-based BCIs.
• To guarantee the recovery of true class means in BCI data.
• To modify the ERP paradigm to integrate seamlessly with machine learning decoders.

Main Methods:

• Introduced Learning from Label Proportions (LLP) as an unsupervised classification approach for ERP-BCIs.
• Developed a visual ERP speller system tailored for LLP requirements.
• Evaluated LLP through simulations on artificial datasets and an online BCI study with 13 participants.

Main Results:

• LLP is theoretically guaranteed to minimize the loss function, similar to supervised classifiers.
• Simulations and online study demonstrated strong performance of LLP.
• An average of 84.5% character spelling accuracy was achieved in the online study without prior calibration.

Conclusions:

• LLP is the first unsupervised decoder for ERP BCIs with a guarantee to find the optimal decoder.
• LLP effectively eliminates the need for tedious calibration sessions.
• LLP's complementary principles offer potential enhancements when combined with existing unsupervised BCI methods.